Usage

pytest                    # unchanged pytest; installing changes nothing
pytest --receptor=llm     # compact, for a coding agent
pytest --receptor=ci      # compact, nothing held back

Full option list in Reference. What it does to pytest underneath, in How it works. What it does not do, in Limitations.

Profiles

	human (default)	llm	ci
Output	unchanged pytest	compact	compact
Plugin registered	none at all	yes	yes
Root causes expanded	—	all, unless >10	all
On-disk report referenced	—	yes	no
Progress on stderr	—	yes	yes

human registers nothing, so plain pytest is byte-identical to not having the receptor installed. Safe to add to a shared environment.

Choosing between llm and ci

One question decides it: will the reader be able to open the on-disk report?

• llm assumes yes — the agent shares the filesystem with the run. So when failures fan out past ~10 distinct root causes, it shows the first ten in full and points at the report for the rest; the agent opens it only if that pathological case actually happens. The trade-off is measured: holding back saves a few tokens per extra cause but costs far more if the file has to be read, so it triggers only on a genuine spread, never on an ordinary handful of failures.

• ci assumes no — a CI runner is destroyed at job end and the log gets one shot, so the report is unreachable by the time anyone reads it. Nothing may be held back and no report path is printed: every root cause is expanded inline.

Same renderer, same truth-preserving rules. They differ only in what survives for the reader — which is why each fits its own context and neither is simply “more” than the other.

What a failing run looks like

FAIL exit=1 | 38 errors, 90 passed | 12.40s | 1 root cause

[1] TypeError | 38 tests | setup
    conftest.py:31
    TypeError: 'NoneType' object is not subscriptable
    tests:
      tests/test_merge.py::test_merge[0]
      tests/test_merge.py::test_merge[1]
      tests/test_merge.py::test_merge[2]
      +35 more
    rerun: pytest tests/test_merge.py -q

• One line on success, with exit status and counts.

• Failures grouped by root cause, keeping every affected test ID.

• No source echo — you have the files. The assertion diff, which you cannot reconstruct, is kept.

• A rerun command that works pasted verbatim from where you invoked pytest.

• Every warning group and every skip reason listed, not counted.

Field-by-field breakdown in Reference.

Flags

You need none. --receptor=llm already quietens pytest further than -q does; adding -q --no-header produces byte-identical output.

Warning

Do not pass --tb=line or --tb=no. They control how pytest builds the traceback, not how it prints it, so they save about twenty tokens and delete the frames: line telling you where to look. Drop a restrictive --tb from addopts for these runs.

Progress

On stderr, after a silent twenty-second warm-up, one line as the run crosses each twenty-percent threshold, ending at 100%:

receptor: 20% 106/530 20s
receptor: 40% 212/530 22s
receptor: 60% 318/530 29s
receptor: 80% 424/530 35s
receptor: 100% 530/530 67s

Thresholds already passed during the warm-up are skipped. From then on, each line is emitted at a live crossing and its percentage is calculated from the count beside it, so the two cannot disagree.

Property
Never on stdout	discard stderr and the report is unchanged
Bounded	at most five lines, whether the run takes five minutes or three hours
Shows pace	a step suddenly taking four times longer is visible
Silent when short	nothing under twenty seconds
Not a hang detector	it fires when a test finishes; a stuck test emits nothing

Distributed runs

pytest --receptor=llm -n 12

Byte-identical output to a serial run. Counts, grouping and exit status do not change with -n, and progress comes from the controller only.

Measuring your own suite

pytest --receptor=llm --receptor-stats

receptor stats: 38 tokens vs 148 for pytest as you configured it | 110 fewer (-74.3%)

The baseline is your configuration, untouched, measured in the same run rather than estimated. Expect it to differ from the published benchmarks, which use a deliberately strict baseline.

Exact counts need tiktoken; without it the figure falls back to a labelled approximation.

Recovering what was held back

Everything you need is on stdout. Detail is withheld only above ten distinct root causes, and only when the report exists to hold it:

full report: .pytest_cache/d/receptor/last-run.txt

Written once, at the end of the run, so recovering it is a file read rather than a second test run. While a run is still going the path does not exist: a previous run’s report is cleared at the start rather than left to be mistaken for the live one, so a mid-run read is an unambiguous “not yet”. With -p no:cacheprovider nothing is withheld at all, since there would be nowhere to recover it from.

--receptor-full expands everything on stdout. It is not the same as --receptor=human: still grouped by root cause, still no source echo.